LLVMBuildExtractElement(builder, descriptor,
LLVMConstInt(ctx->i32, 2, 0), "");
- if (ctx->screen->info.chip_class == VI) {
- /* On VI, the descriptor contains the size in bytes,
+ if (ctx->screen->info.chip_class == GFX8) {
+ /* On GFX8, the descriptor contains the size in bytes,
* but TXQ must return the size in elements.
* The stride is always non-zero for resources using TXQ.
*/
switch (target) {
case TGSI_TEXTURE_1D:
case TGSI_TEXTURE_SHADOW1D:
- if (screen->info.chip_class >= GFX9)
+ if (screen->info.chip_class == GFX9)
return ac_image_2d;
return ac_image_1d;
case TGSI_TEXTURE_2D:
return ac_image_cube;
case TGSI_TEXTURE_1D_ARRAY:
case TGSI_TEXTURE_SHADOW1D_ARRAY:
- if (screen->info.chip_class >= GFX9)
+ if (screen->info.chip_class == GFX9)
return ac_image_2darray;
return ac_image_1darray;
case TGSI_TEXTURE_2D_ARRAY:
/* Match the resource type set in the descriptor. */
if (dim == ac_image_cube ||
- (screen->info.chip_class <= VI && dim == ac_image_3d))
+ (screen->info.chip_class <= GFX8 && dim == ac_image_3d))
dim = ac_image_2darray;
- else if (target == TGSI_TEXTURE_2D && screen->info.chip_class >= GFX9) {
+ else if (target == TGSI_TEXTURE_2D && screen->info.chip_class == GFX9) {
/* When a single layer of a 3D texture is bound, the shader
* will refer to a 2D target, but the descriptor has a 3D type.
* Since the HW ignores BASE_ARRAY in this case, we need to
static LLVMValueRef force_dcc_off(struct si_shader_context *ctx,
LLVMValueRef rsrc)
{
- if (ctx->screen->info.chip_class <= CIK) {
+ if (ctx->screen->info.chip_class <= GFX7) {
return rsrc;
} else {
LLVMValueRef i32_6 = LLVMConstInt(ctx->i32, 6, 0);
LLVMValueRef si_load_image_desc(struct si_shader_context *ctx,
LLVMValueRef list, LLVMValueRef index,
- enum ac_descriptor_type desc_type, bool dcc_off,
- bool bindless)
+ enum ac_descriptor_type desc_type,
+ bool uses_store, bool bindless)
{
LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef rsrc;
else
rsrc = ac_build_load_to_sgpr(&ctx->ac, list, index);
- if (desc_type == AC_DESC_IMAGE && dcc_off)
+ if (desc_type == AC_DESC_IMAGE && uses_store)
rsrc = force_dcc_off(ctx, rsrc);
return rsrc;
}
LLVMValueRef rsrc_ptr = LLVMGetParam(ctx->main_fn,
ctx->param_samplers_and_images);
LLVMValueRef index;
- bool dcc_off = is_store;
if (!image->Register.Indirect) {
- const struct tgsi_shader_info *info = bld_base->info;
- unsigned images_writemask = info->images_store |
- info->images_atomic;
-
index = LLVMConstInt(ctx->i32,
si_get_image_slot(image->Register.Index), 0);
-
- if (images_writemask & (1 << image->Register.Index))
- dcc_off = true;
} else {
/* From the GL_ARB_shader_image_load_store extension spec:
*
*rsrc = si_load_image_desc(ctx, rsrc_ptr, index,
target == TGSI_TEXTURE_BUFFER ? AC_DESC_BUFFER : AC_DESC_IMAGE,
- dcc_off, bindless);
+ is_store, bindless);
}
static void image_fetch_coords(
coords[chan] = tmp;
}
- if (ctx->screen->info.chip_class >= GFX9) {
+ if (ctx->screen->info.chip_class == GFX9) {
/* 1D textures are allocated and used as 2D on GFX9. */
if (target == TGSI_TEXTURE_1D) {
coords[1] = ctx->i32_0;
unsigned cache_policy = 0;
if (!atomic &&
- /* SI has a TC L1 bug causing corruption of 8bit/16bit stores.
+ /* GFX6 has a TC L1 bug causing corruption of 8bit/16bit stores.
* All store opcodes not aligned to a dword are affected.
* The only way to get unaligned stores in radeonsi is through
* shader images. */
- ((may_store_unaligned && ctx->screen->info.chip_class == SI) ||
+ ((may_store_unaligned && ctx->screen->info.chip_class == GFX6) ||
/* If this is write-only, don't keep data in L1 to prevent
* evicting L1 cache lines that may be needed by other
* instructions. */
writeonly_memory ||
- inst->Memory.Qualifier & (TGSI_MEMORY_COHERENT | TGSI_MEMORY_VOLATILE)))
+ inst->Memory.Qualifier & (TGSI_MEMORY_COHERENT | TGSI_MEMORY_VOLATILE))) {
cache_policy |= ac_glc;
+ }
if (inst->Memory.Qualifier & TGSI_MEMORY_STREAM_CACHE_POLICY)
cache_policy |= ac_slc;
* For LOAD, set this to (store | atomic) slot usage in the shader.
* For STORE, set this to (load | atomic) slot usage in the shader.
* \param images_reverse_access_mask Same as above, but for images.
+ * \param bindless_buffer_reverse_access_mask Same as above, but for bindless image buffers.
+ * \param bindless_image_reverse_access_mask Same as above, but for bindless images.
*/
static bool is_oneway_access_only(const struct tgsi_full_instruction *inst,
const struct tgsi_shader_info *info,
unsigned shader_buffers_reverse_access_mask,
- unsigned images_reverse_access_mask)
+ unsigned images_reverse_access_mask,
+ bool bindless_buffer_reverse_access_mask,
+ bool bindless_image_reverse_access_mask)
{
+ enum tgsi_file_type resource_file;
+ unsigned resource_index;
+ bool resource_indirect;
+
+ if (inst->Instruction.Opcode == TGSI_OPCODE_STORE) {
+ resource_file = inst->Dst[0].Register.File;
+ resource_index = inst->Dst[0].Register.Index;
+ resource_indirect = inst->Dst[0].Register.Indirect;
+ } else {
+ resource_file = inst->Src[0].Register.File;
+ resource_index = inst->Src[0].Register.Index;
+ resource_indirect = inst->Src[0].Register.Indirect;
+ }
+
+ assert(resource_file == TGSI_FILE_BUFFER ||
+ resource_file == TGSI_FILE_IMAGE ||
+ /* bindless image */
+ resource_file == TGSI_FILE_INPUT ||
+ resource_file == TGSI_FILE_OUTPUT ||
+ resource_file == TGSI_FILE_CONSTANT ||
+ resource_file == TGSI_FILE_TEMPORARY ||
+ resource_file == TGSI_FILE_IMMEDIATE);
+
+ assert(resource_file != TGSI_FILE_BUFFER ||
+ inst->Memory.Texture == TGSI_TEXTURE_BUFFER);
+
+ bool bindless = resource_file != TGSI_FILE_BUFFER &&
+ resource_file != TGSI_FILE_IMAGE;
+
/* RESTRICT means NOALIAS.
* If there are no writes, we can assume the accessed memory is read-only.
* If there are no reads, we can assume the accessed memory is write-only.
*/
- if (inst->Memory.Qualifier & TGSI_MEMORY_RESTRICT) {
+ if (inst->Memory.Qualifier & TGSI_MEMORY_RESTRICT && !bindless) {
unsigned reverse_access_mask;
- if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
+ if (resource_file == TGSI_FILE_BUFFER) {
reverse_access_mask = shader_buffers_reverse_access_mask;
} else if (inst->Memory.Texture == TGSI_TEXTURE_BUFFER) {
reverse_access_mask = info->images_buffers &
images_reverse_access_mask;
}
- if (inst->Src[0].Register.Indirect) {
+ if (resource_indirect) {
if (!reverse_access_mask)
return true;
} else {
if (!(reverse_access_mask &
- (1u << inst->Src[0].Register.Index)))
+ (1u << resource_index)))
return true;
}
}
* Same for the case when there are no writes/reads for non-buffer
* images.
*/
- if (inst->Src[0].Register.File == TGSI_FILE_BUFFER ||
- (inst->Memory.Texture == TGSI_TEXTURE_BUFFER &&
- (inst->Src[0].Register.File == TGSI_FILE_IMAGE ||
- tgsi_is_bindless_image_file(inst->Src[0].Register.File)))) {
+ if (resource_file == TGSI_FILE_BUFFER ||
+ inst->Memory.Texture == TGSI_TEXTURE_BUFFER) {
if (!shader_buffers_reverse_access_mask &&
- !(info->images_buffers & images_reverse_access_mask))
+ !(info->images_buffers & images_reverse_access_mask) &&
+ !bindless_buffer_reverse_access_mask)
return true;
} else {
- if (!(~info->images_buffers & images_reverse_access_mask))
+ if (!(~info->images_buffers & images_reverse_access_mask) &&
+ !bindless_image_reverse_access_mask)
return true;
}
return false;
bool ubo = inst->Src[0].Register.File == TGSI_FILE_CONSTBUF;
args.resource = shader_buffer_fetch_rsrc(ctx, &inst->Src[0], ubo);
voffset = ac_to_integer(&ctx->ac, lp_build_emit_fetch(bld_base, inst, 1, 0));
- } else if (inst->Src[0].Register.File == TGSI_FILE_IMAGE ||
- tgsi_is_bindless_image_file(inst->Src[0].Register.File)) {
+ } else {
unsigned target = inst->Memory.Texture;
image_fetch_rsrc(bld_base, &inst->Src[0], false, target, &args.resource);
emit_data->output[emit_data->chan] =
ac_build_buffer_load(&ctx->ac, args.resource,
util_last_bit(inst->Dst[0].Register.WriteMask),
- NULL, voffset, NULL, 0, 0, 0, true, true);
+ NULL, voffset, NULL, 0, 0, true, true);
return;
}
if (inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE)
- ac_build_waitcnt(&ctx->ac, VM_CNT);
+ ac_build_waitcnt(&ctx->ac, AC_WAIT_VLOAD | AC_WAIT_VSTORE);
can_speculate = !(inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE) &&
is_oneway_access_only(inst, info,
info->shader_buffers_store |
info->shader_buffers_atomic,
info->images_store |
- info->images_atomic);
+ info->images_atomic,
+ info->uses_bindless_buffer_store |
+ info->uses_bindless_buffer_atomic,
+ info->uses_bindless_image_store |
+ info->uses_bindless_image_atomic);
args.cache_policy = get_cache_policy(ctx, inst, false, false, false);
if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
ac_build_buffer_load(&ctx->ac, args.resource,
util_last_bit(inst->Dst[0].Register.WriteMask),
NULL, voffset, NULL, 0,
- !!(args.cache_policy & ac_glc),
- !!(args.cache_policy & ac_slc),
- can_speculate, false);
+ args.cache_policy, can_speculate, false);
return;
}
vindex,
ctx->i32_0,
num_channels,
- !!(args.cache_policy & ac_glc),
+ args.cache_policy,
can_speculate);
emit_data->output[emit_data->chan] =
ac_build_expand_to_vec4(&ctx->ac, result, num_channels);
while (writemask) {
int start, count;
- const char *intrinsic_name;
LLVMValueRef data, voff;
u_bit_scan_consecutive_range(&writemask, &start, &count);
- /* Due to an LLVM limitation, split 3-element writes
- * into a 2-element and a 1-element write. */
- if (count == 3) {
- writemask |= 1 << (start + 2);
- count = 2;
- }
-
- if (count == 4) {
+ if (count == 3 && ac_has_vec3_support(ctx->ac.chip_class, false)) {
+ LLVMValueRef values[3] = {
+ LLVMBuildExtractElement(builder, base_data,
+ LLVMConstInt(ctx->i32, start, 0), ""),
+ LLVMBuildExtractElement(builder, base_data,
+ LLVMConstInt(ctx->i32, start + 1, 0), ""),
+ LLVMBuildExtractElement(builder, base_data,
+ LLVMConstInt(ctx->i32, start + 2, 0), ""),
+ };
+ data = ac_build_gather_values(&ctx->ac, values, 3);
+ } else if (count >= 3) {
data = base_data;
- intrinsic_name = "llvm.amdgcn.buffer.store.v4f32";
} else if (count == 2) {
LLVMValueRef values[2] = {
LLVMBuildExtractElement(builder, base_data,
};
data = ac_build_gather_values(&ctx->ac, values, 2);
- intrinsic_name = "llvm.amdgcn.buffer.store.v2f32";
} else {
assert(count == 1);
data = LLVMBuildExtractElement(
builder, base_data,
LLVMConstInt(ctx->i32, start, 0), "");
- intrinsic_name = "llvm.amdgcn.buffer.store.f32";
}
voff = base_offset;
LLVMConstInt(ctx->i32, start * 4, 0), "");
}
- LLVMValueRef args[] = {
- data,
- resource,
- ctx->i32_0, /* vindex */
- voff,
- LLVMConstInt(ctx->i1, !!(cache_policy & ac_glc), 0),
- LLVMConstInt(ctx->i1, !!(cache_policy & ac_slc), 0),
- };
- ac_build_intrinsic(&ctx->ac, intrinsic_name, ctx->voidt, args, 6,
- ac_get_store_intr_attribs(writeonly_memory));
+ ac_build_buffer_store_dword(&ctx->ac, resource, data, count,
+ voff, ctx->i32_0, 0, cache_policy,
+ false);
}
}
struct tgsi_full_src_register resource_reg =
tgsi_full_src_register_from_dst(&inst->Dst[0]);
unsigned target = inst->Memory.Texture;
+
+ if (inst->Dst[0].Register.File == TGSI_FILE_MEMORY) {
+ store_emit_memory(ctx, emit_data);
+ return;
+ }
+
bool writeonly_memory = is_oneway_access_only(inst, info,
info->shader_buffers_load |
info->shader_buffers_atomic,
info->images_load |
- info->images_atomic);
- bool is_image = inst->Dst[0].Register.File == TGSI_FILE_IMAGE ||
- tgsi_is_bindless_image_file(inst->Dst[0].Register.File);
- LLVMValueRef chans[4], value;
+ info->images_atomic,
+ info->uses_bindless_buffer_load |
+ info->uses_bindless_buffer_atomic,
+ info->uses_bindless_image_load |
+ info->uses_bindless_image_atomic);
+ LLVMValueRef chans[4];
LLVMValueRef vindex = ctx->i32_0;
LLVMValueRef voffset = ctx->i32_0;
struct ac_image_args args = {};
- if (inst->Dst[0].Register.File == TGSI_FILE_MEMORY) {
- store_emit_memory(ctx, emit_data);
- return;
- }
-
for (unsigned chan = 0; chan < 4; ++chan)
chans[chan] = lp_build_emit_fetch(bld_base, inst, 1, chan);
- value = ac_build_gather_values(&ctx->ac, chans, 4);
-
if (inst->Dst[0].Register.File == TGSI_FILE_BUFFER) {
args.resource = shader_buffer_fetch_rsrc(ctx, &resource_reg, false);
voffset = ac_to_integer(&ctx->ac, lp_build_emit_fetch(bld_base, inst, 0, 0));
- } else if (is_image) {
+ } else {
image_fetch_rsrc(bld_base, &resource_reg, true, target, &args.resource);
image_fetch_coords(bld_base, inst, 0, args.resource, args.coords);
vindex = args.coords[0]; /* for buffers only */
- } else {
- unreachable("unexpected register file");
}
if (inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE)
- ac_build_waitcnt(&ctx->ac, VM_CNT);
+ ac_build_waitcnt(&ctx->ac, AC_WAIT_VLOAD | AC_WAIT_VSTORE);
+ bool is_image = inst->Dst[0].Register.File != TGSI_FILE_BUFFER;
args.cache_policy = get_cache_policy(ctx, inst,
false, /* atomic */
is_image, /* may_store_unaligned */
if (inst->Dst[0].Register.File == TGSI_FILE_BUFFER) {
store_emit_buffer(ctx, args.resource, inst->Dst[0].Register.WriteMask,
- value, voffset, args.cache_policy, writeonly_memory);
+ ac_build_gather_values(&ctx->ac, chans, 4),
+ voffset, args.cache_policy, writeonly_memory);
return;
}
if (target == TGSI_TEXTURE_BUFFER) {
- LLVMValueRef buf_args[] = {
- value,
- args.resource,
- vindex,
- ctx->i32_0, /* voffset */
- LLVMConstInt(ctx->i1, !!(args.cache_policy & ac_glc), 0),
- LLVMConstInt(ctx->i1, !!(args.cache_policy & ac_slc), 0),
- };
-
- emit_data->output[emit_data->chan] = ac_build_intrinsic(
- &ctx->ac, "llvm.amdgcn.buffer.store.format.v4f32",
- ctx->voidt, buf_args, 6,
- ac_get_store_intr_attribs(writeonly_memory));
+ unsigned num_channels = util_last_bit(inst->Dst[0].Register.WriteMask);
+
+ ac_build_buffer_store_format(&ctx->ac, args.resource,
+ ac_build_gather_values(&ctx->ac, chans, num_channels),
+ vindex, ctx->i32_0 /* voffset */,
+ num_channels,
+ args.cache_policy);
} else {
args.opcode = ac_image_store;
- args.data[0] = value;
+ args.data[0] = ac_build_gather_values(&ctx->ac, chans, 4);
args.dim = ac_image_dim_from_tgsi_target(ctx->screen, inst->Memory.Texture);
- args.attributes = ac_get_store_intr_attribs(writeonly_memory);
+ args.attributes = AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY;
args.dmask = 0xf;
emit_data->output[emit_data->chan] =
LLVMBuilderRef builder = ctx->ac.builder;
const struct tgsi_full_instruction * inst = emit_data->inst;
LLVMValueRef ptr, result, arg;
+ const char *sync_scope = HAVE_LLVM >= 0x0900 ? "workgroup-one-as" : "workgroup";
ptr = get_memory_ptr(ctx, inst, ctx->i32, 1);
new_data = ac_to_integer(&ctx->ac, new_data);
- result = LLVMBuildAtomicCmpXchg(builder, ptr, arg, new_data,
- LLVMAtomicOrderingSequentiallyConsistent,
- LLVMAtomicOrderingSequentiallyConsistent,
- false);
-
+ result = ac_build_atomic_cmp_xchg(&ctx->ac, ptr, arg, new_data,
+ sync_scope);
result = LLVMBuildExtractValue(builder, result, 0, "");
} else {
LLVMAtomicRMWBinOp op;
unreachable("unknown atomic opcode");
}
- result = LLVMBuildAtomicRMW(builder, op, ptr, arg,
- LLVMAtomicOrderingSequentiallyConsistent,
- false);
+ result = ac_build_atomic_rmw(&ctx->ac, op, ptr, arg, sync_scope);
}
emit_data->output[emit_data->chan] =
LLVMBuildBitCast(builder, result, ctx->f32, "");
args.data[num_data++] =
ac_to_integer(&ctx->ac, lp_build_emit_fetch(bld_base, inst, 2, 0));
+
args.cache_policy = get_cache_policy(ctx, inst, true, false, false);
if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
args.resource = shader_buffer_fetch_rsrc(ctx, &inst->Src[0], false);
voffset = ac_to_integer(&ctx->ac, lp_build_emit_fetch(bld_base, inst, 1, 0));
- } else if (inst->Src[0].Register.File == TGSI_FILE_IMAGE ||
- tgsi_is_bindless_image_file(inst->Src[0].Register.File)) {
+ } else {
image_fetch_rsrc(bld_base, &inst->Src[0], true,
inst->Memory.Texture, &args.resource);
image_fetch_coords(bld_base, inst, 1, args.resource, args.coords);
vindex = args.coords[0]; /* for buffers only */
}
- if (inst->Src[0].Register.File == TGSI_FILE_BUFFER ||
+ if (HAVE_LLVM >= 0x0800 &&
+ inst->Src[0].Register.File != TGSI_FILE_BUFFER &&
inst->Memory.Texture == TGSI_TEXTURE_BUFFER) {
LLVMValueRef buf_args[7];
unsigned num_args = 0;
+ buf_args[num_args++] = args.data[0];
+ if (inst->Instruction.Opcode == TGSI_OPCODE_ATOMCAS)
+ buf_args[num_args++] = args.data[1];
+
+ buf_args[num_args++] = args.resource;
+ buf_args[num_args++] = vindex;
+ buf_args[num_args++] = voffset;
+ buf_args[num_args++] = ctx->i32_0; /* soffset */
+ buf_args[num_args++] = LLVMConstInt(ctx->i32, args.cache_policy & ac_slc, 0);
+
+ char intrinsic_name[64];
+ snprintf(intrinsic_name, sizeof(intrinsic_name),
+ "llvm.amdgcn.struct.buffer.atomic.%s", action->intr_name);
+ emit_data->output[emit_data->chan] =
+ ac_to_float(&ctx->ac,
+ ac_build_intrinsic(&ctx->ac, intrinsic_name,
+ ctx->i32, buf_args, num_args, 0));
+ return;
+ }
+
+ if (inst->Src[0].Register.File == TGSI_FILE_BUFFER ||
+ (HAVE_LLVM < 0x0800 &&
+ inst->Memory.Texture == TGSI_TEXTURE_BUFFER)) {
+ LLVMValueRef buf_args[7];
+ unsigned num_args = 0;
+
buf_args[num_args++] = args.data[0];
if (inst->Instruction.Opcode == TGSI_OPCODE_ATOMCAS)
buf_args[num_args++] = args.data[1];
case TGSI_OPCODE_ATOMUMAX: args.atomic = ac_atomic_umax; break;
case TGSI_OPCODE_ATOMIMIN: args.atomic = ac_atomic_smin; break;
case TGSI_OPCODE_ATOMIMAX: args.atomic = ac_atomic_smax; break;
+ case TGSI_OPCODE_ATOMINC_WRAP:
+ args.atomic = ac_atomic_inc_wrap;
+ break;
+ case TGSI_OPCODE_ATOMDEC_WRAP:
+ args.atomic = ac_atomic_dec_wrap;
+ break;
default: unreachable("unhandled image atomic");
}
}
LLVMBuilderRef builder = ctx->ac.builder;
/* 1D textures are allocated and used as 2D on GFX9. */
- if (ctx->screen->info.chip_class >= GFX9 &&
+ if (ctx->screen->info.chip_class == GFX9 &&
(target == TGSI_TEXTURE_1D_ARRAY ||
target == TGSI_TEXTURE_SHADOW1D_ARRAY)) {
LLVMValueRef layers =
args.opcode = ac_image_get_resinfo;
args.dim = ac_texture_dim_from_tgsi_target(ctx->screen, target);
args.dmask = 0xf;
+ args.attributes = AC_FUNC_ATTR_READNONE;
if (inst->Instruction.Opcode == TGSI_OPCODE_TXQ) {
tex_fetch_ptrs(bld_base, emit_data, &args.resource, NULL, NULL);
list = LLVMBuildPointerCast(builder, list,
ac_array_in_const32_addr_space(ctx->v4i32), "");
break;
+ case AC_DESC_PLANE_0:
+ case AC_DESC_PLANE_1:
+ case AC_DESC_PLANE_2:
+ /* Only used for the multiplane image support for Vulkan. Should
+ * never be reached in radeonsi.
+ */
+ unreachable("Plane descriptor requested in radeonsi.");
}
return ac_build_load_to_sgpr(&ctx->ac, list, index);
/* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
*
- * SI-CI:
+ * GFX6-GFX7:
* If BASE_LEVEL == LAST_LEVEL, the shader must disable anisotropic
* filtering manually. The driver sets img7 to a mask clearing
* MAX_ANISO_RATIO if BASE_LEVEL == LAST_LEVEL. The shader must do:
* s_and_b32 samp0, samp0, img7
*
- * VI:
+ * GFX8:
* The ANISO_OVERRIDE sampler field enables this fix in TA.
*/
static LLVMValueRef sici_fix_sampler_aniso(struct si_shader_context *ctx,
{
LLVMValueRef img7, samp0;
- if (ctx->screen->info.chip_class >= VI)
+ if (ctx->screen->info.chip_class >= GFX8)
return samp;
img7 = LLVMBuildExtractElement(ctx->ac.builder, res,
uint32_t wa_num_format =
return_type == TGSI_RETURN_TYPE_UINT ?
- S_008F14_NUM_FORMAT_GFX6(V_008F14_IMG_NUM_FORMAT_USCALED) :
- S_008F14_NUM_FORMAT_GFX6(V_008F14_IMG_NUM_FORMAT_SSCALED);
+ S_008F14_NUM_FORMAT(V_008F14_IMG_NUM_FORMAT_USCALED) :
+ S_008F14_NUM_FORMAT(V_008F14_IMG_NUM_FORMAT_SSCALED);
wa_formats = LLVMBuildAnd(builder, formats,
- LLVMConstInt(ctx->i32, C_008F14_NUM_FORMAT_GFX6, false),
+ LLVMConstInt(ctx->i32, C_008F14_NUM_FORMAT, false),
"");
wa_formats = LLVMBuildOr(builder, wa_formats,
LLVMConstInt(ctx->i32, wa_num_format, false), "");
resinfo.sampler = args->sampler;
resinfo.lod = ctx->ac.i32_0;
resinfo.dmask = 0xf;
+ resinfo.attributes = AC_FUNC_ATTR_READNONE;
LLVMValueRef texsize =
fix_resinfo(ctx, target,
args.resource,
vindex,
ctx->i32_0,
- num_channels, false, true);
+ num_channels, 0, true);
emit_data->output[emit_data->chan] =
ac_build_expand_to_vec4(&ctx->ac, result, num_channels);
return;
*
* TC-compatible HTILE promotes Z16 and Z24 to Z32_FLOAT,
* so the depth comparison value isn't clamped for Z16 and
- * Z24 anymore. Do it manually here.
+ * Z24 anymore. Do it manually here for GFX8-9; GFX10 has
+ * an explicitly clamped 32-bit float format.
*/
- if (ctx->screen->info.chip_class >= VI) {
+ if (ctx->screen->info.chip_class >= GFX8 &&
+ ctx->screen->info.chip_class <= GFX9) {
LLVMValueRef upgraded;
LLVMValueRef clamped;
upgraded = LLVMBuildExtractElement(ctx->ac.builder, args.sampler,
num_src_deriv_channels = 1;
/* 1D textures are allocated and used as 2D on GFX9. */
- if (ctx->screen->info.chip_class >= GFX9) {
+ if (ctx->screen->info.chip_class == GFX9) {
num_dst_deriv_channels = 2;
} else {
num_dst_deriv_channels = 1;
} else if (tgsi_is_array_sampler(target) &&
opcode != TGSI_OPCODE_TXF &&
opcode != TGSI_OPCODE_TXF_LZ &&
- ctx->screen->info.chip_class <= VI) {
+ ctx->screen->info.chip_class <= GFX8) {
unsigned array_coord = target == TGSI_TEXTURE_1D_ARRAY ? 1 : 2;
args.coords[array_coord] = ac_build_round(&ctx->ac, args.coords[array_coord]);
}
/* 1D textures are allocated and used as 2D on GFX9. */
- if (ctx->screen->info.chip_class >= GFX9) {
+ if (ctx->screen->info.chip_class == GFX9) {
LLVMValueRef filler;
/* Use 0.5, so that we don't sample the border color. */
}
}
- if (target == TGSI_TEXTURE_2D_MSAA ||
- target == TGSI_TEXTURE_2D_ARRAY_MSAA) {
+ if ((target == TGSI_TEXTURE_2D_MSAA ||
+ target == TGSI_TEXTURE_2D_ARRAY_MSAA) &&
+ !(ctx->screen->debug_flags & DBG(NO_FMASK))) {
ac_apply_fmask_to_sample(&ctx->ac, fmask_ptr, args.coords,
target == TGSI_TEXTURE_2D_ARRAY_MSAA);
}
/* The hardware needs special lowering for Gather4 with integer formats. */
LLVMValueRef gather4_int_result_workaround = NULL;
- if (ctx->screen->info.chip_class <= VI &&
+ if (ctx->screen->info.chip_class <= GFX8 &&
opcode == TGSI_OPCODE_TG4) {
assert(inst->Texture.ReturnType != TGSI_RETURN_TYPE_UNKNOWN);
emit_data->output[emit_data->chan] = samples;
}
-static void si_llvm_emit_fbfetch(const struct lp_build_tgsi_action *action,
- struct lp_build_tgsi_context *bld_base,
- struct lp_build_emit_data *emit_data)
+static LLVMValueRef si_llvm_emit_fbfetch(struct si_shader_context *ctx)
{
- struct si_shader_context *ctx = si_shader_context(bld_base);
struct ac_image_args args = {};
LLVMValueRef ptr, image, fmask;
if (ctx->shader->key.mono.u.ps.fbfetch_msaa)
args.coords[chan++] = si_get_sample_id(ctx);
- if (ctx->shader->key.mono.u.ps.fbfetch_msaa) {
+ if (ctx->shader->key.mono.u.ps.fbfetch_msaa &&
+ !(ctx->screen->debug_flags & DBG(NO_FMASK))) {
fmask = ac_build_load_to_sgpr(&ctx->ac, ptr,
LLVMConstInt(ctx->i32, SI_PS_IMAGE_COLORBUF0_FMASK / 2, 0));
args.opcode = ac_image_load;
args.resource = image;
args.dmask = 0xf;
+ args.attributes = AC_FUNC_ATTR_READNONE;
+
if (ctx->shader->key.mono.u.ps.fbfetch_msaa)
args.dim = ctx->shader->key.mono.u.ps.fbfetch_layered ?
ac_image_2darraymsaa : ac_image_2dmsaa;
args.dim = ctx->shader->key.mono.u.ps.fbfetch_layered ?
ac_image_2darray : ac_image_2d;
- emit_data->output[emit_data->chan] =
- ac_build_image_opcode(&ctx->ac, &args);
+ return ac_build_image_opcode(&ctx->ac, &args);
+}
+
+static void si_tgsi_emit_fbfetch(const struct lp_build_tgsi_action *action,
+ struct lp_build_tgsi_context *bld_base,
+ struct lp_build_emit_data *emit_data)
+{
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+
+ emit_data->output[emit_data->chan] = si_llvm_emit_fbfetch(ctx);
+}
+
+LLVMValueRef si_nir_emit_fbfetch(struct ac_shader_abi *abi)
+{
+ struct si_shader_context *ctx = si_shader_context_from_abi(abi);
+
+ return si_llvm_emit_fbfetch(ctx);
}
/**
bld_base->op_actions[TGSI_OPCODE_LODQ].emit = build_tex_intrinsic;
bld_base->op_actions[TGSI_OPCODE_TXQS].emit = si_llvm_emit_txqs;
- bld_base->op_actions[TGSI_OPCODE_FBFETCH].emit = si_llvm_emit_fbfetch;
+ bld_base->op_actions[TGSI_OPCODE_FBFETCH].emit = si_tgsi_emit_fbfetch;
bld_base->op_actions[TGSI_OPCODE_LOAD].emit = load_emit;
bld_base->op_actions[TGSI_OPCODE_STORE].emit = store_emit;
bld_base->op_actions[TGSI_OPCODE_ATOMIMIN].intr_name = "smin";
bld_base->op_actions[TGSI_OPCODE_ATOMIMAX].emit = atomic_emit;
bld_base->op_actions[TGSI_OPCODE_ATOMIMAX].intr_name = "smax";
+ bld_base->op_actions[TGSI_OPCODE_ATOMINC_WRAP].emit = atomic_emit;
+ bld_base->op_actions[TGSI_OPCODE_ATOMINC_WRAP].intr_name = "inc";
+ bld_base->op_actions[TGSI_OPCODE_ATOMDEC_WRAP].emit = atomic_emit;
+ bld_base->op_actions[TGSI_OPCODE_ATOMDEC_WRAP].intr_name = "dec";
}